Bi-Mamba-Based Net-Load Forecasting Method with Multidimensional Temporal Information Fusion
TL;DR
A net-load prediction method considering multidimensional timing information is proposed, which can reduce the normalized Mean Absolute Error and the normalized Root Mean Squared Error compared with the temporal convolutional network baseline and exhibits robust stability across different seasons and day types.
Abstract
With the rise in small-scale distributed photovoltaic (PV) power generation technology, the behind-the-meter PV problem has greatly increased the difficulty of power system regulation and management and accurate net-load forecasting is of great significance to the economic and stable operation of the power system. The timing features of the net-load sequence are complex due to a variety of factors. In order to improve the extraction effect of the timing model on the timing features of the net-load sequence and to increase the accuracy of the net-load prediction, a net-load prediction method considering multidimensional timing information is proposed. A Mamba module is introduced into the model to filter the input data, retaining some of the effective contextual information while improving the operational efficiency of the model. The structure of Bi-Mamba is used to construct a bidirectional time-series feature extraction model, which fuses the date attributes and the positive and negative time-series features of the net load to improve the stability and accuracy of the model prediction. The results of the validation algorithms show that the proposed method can reduce the normalized Mean Absolute Error (nMAE) by 17.72% and the normalized Root Mean Squared Error (nRMSE) by 21.51% compared with the temporal convolutional network (TCN) baseline. Furthermore, the model exhibits robust stability across different seasons and day types, providing a reliable reference for scheduling decisions in power systems with high PV penetration.